Safety Evaluation on CategoricalHarmfulQA Dolly fine-tuning (test)
1.27ASR (S1)LARF
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LARFBase Model=Llama2-7b-Chat2026.05 | 1.27 | 3.45 | 15.82 | 6.55 | 1.64 | -0.37 | — | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Qwen3-8B2026.05 | 3.45 | 28.91 | 38.73 | 22.36 | 15.64 | -12.19 | — | — | — | — | — | — | — | |
| Self-Inf-NBase Model=Llama2-7b-Chat2026.05 | 3.45 | 4.55 | 0.91 | 7.09 | 1.64 | 1.81 | — | — | — | — | — | — | — | |
| Bi-Anchor(Reps)Base Model=Qwen3-8B2026.05 | 9.64 | 16.36 | 52.18 | 27.64 | 25.64 | -16 | — | — | — | — | — | — | — | |
| Bi-Anchor(Reps)Base Model=Llama2-7b-Chat2026.05 | 13.27 | 10.36 | 5.82 | 0.55 | 10.55 | 2.72 | — | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Llama2-7b-Chat2026.05 | 22.36 | 4.36 | 4.73 | 2 | 11.82 | 10.54 | — | — | — | — | — | — | — | |
| Self-Inf-NBase Model=Llama3.1-8B-Instruct2026.05 | 22.91 | 38.73 | 27.64 | 27.64 | 27.64 | -4.73 | — | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Llama2-7b-Chat2026.05 | 43.27 | 17.27 | 8.36 | 3.27 | 0 | 43.27 | — | — | — | — | — | — | — | |
| Reward ModelBase Model=Llama2-7b-Chat2026.05 | 44.36 | 6 | 6 | 1.64 | 0 | 44.36 | — | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Qwen3-8B2026.05 | 45.27 | 28.36 | 15.45 | 4.55 | 2.73 | 42.54 | — | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Llama2-7b-Chat2026.05 | 45.27 | 14.73 | 5.45 | 2.55 | 0.36 | 44.91 | — | — | — | — | — | — | — | |
| Bi-Anchor(Reps)Base Model=Llama3.1-8B-Instruct2026.05 | 46.36 | 37.09 | 40.55 | 42 | 23.09 | 23.27 | — | — | — | — | — | — | — | |
| Self-Inf-NBase Model=Qwen3-8B2026.05 | 48.36 | 34.73 | 7.27 | 36.36 | 20.91 | 27.45 | — | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Llama3.1-8B-Instruct2026.05 | 56.55 | 50.36 | 35.64 | 22.18 | 16.91 | 39.64 | — | — | — | — | — | — | — | |
| Reward ModelBase Model=Qwen3-8B2026.05 | 59.82 | 14.91 | 4.55 | 13.82 | 5.45 | 54.37 | — | — | — | — | — | — | — | |
| LARFBase Model=Llama3.1-8B-Instruct2026.05 | 61.27 | 68.73 | 33.09 | 43.82 | 11.09 | 50.18 | — | — | — | — | — | — | — | |
| Reward ModelBase Model=Llama3.1-8B-Instruct2026.05 | 67.64 | 43.27 | 18.55 | 37.82 | 20.18 | 47.46 | — | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Qwen3-8B2026.05 | 71.27 | 29.45 | 10.18 | 7.27 | 2.55 | 68.72 | — | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Llama3.1-8B-Instruct2026.05 | 76.36 | 21.45 | 36.36 | 16.18 | 13.82 | 62.54 | — | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Llama3.1-8B-Instruct2026.05 | 79.82 | 57.27 | 37.45 | 5.27 | 4.73 | 75.09 | — | — | — | — | — | — | — | |
| LARFBase Model=Qwen3-8B2026.05 | 84.18 | 70.91 | 12.73 | 18.91 | 1.09 | 83.09 | — | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Backbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | 2.92 | -0.51 | -1.49 | 0.26 | 0.54 | -2.38 | |
| Bi-Anchor(Grad)Backbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -2.39 | -3.04 | -1.79 | -0.6 | 0.57 | 2.96 | |
| Bi-Anchor(Grad)Backbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | 1.98 | 3.42 | 3.29 | 3.96 | 2.64 | 0.66 | |
| Bi-Anchor(Reps)Backbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | 1.02 | 1.12 | -2.3 | 0.04 | 0.21 | -0.81 | |
| Bi-Anchor(Reps)Backbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -1.85 | -1.93 | -2.4 | -2.73 | -0.95 | 0.9 | |
| Bi-Anchor(Reps)Backbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | 2.5 | 2.31 | 3.69 | 5.14 | 2.92 | 0.42 | |
| LARFBackbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | -5.65 | -3.66 | 1.41 | 0.9 | 3.23 | 8.88 | |
| LARFBackbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -4.29 | -4.28 | -1.1 | -1.64 | 0.99 | 5.28 | |
| LARFBackbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | 4.39 | 4.85 | 1.98 | 2.79 | 2.99 | -1.4 | |
| Reward ModelBackbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | -3.44 | 0.59 | 1.95 | 1.52 | 3.65 | 7.09 | |
| Reward ModelBackbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -3.68 | -2.12 | 0.34 | -2.24 | -0.89 | 2.79 | |
| Reward ModelBackbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | -1.21 | 2.83 | 3.37 | 4.39 | 5.66 | 6.87 | |
| Self-Inf-NBackbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | -1.03 | -0.51 | 2.28 | -1.24 | 1.01 | 2.04 | |
| Self-Inf-NBackbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | 0.12 | -1.37 | -3.02 | -3.73 | -1.17 | -1.29 | |
| Self-Inf-NBackbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | 3.95 | 3.98 | 4.24 | 3.8 | 4.5 | 0.55 | |
| SQSD(Aegis)Backbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | -1.78 | 0.03 | 1.36 | 1.81 | 3.08 | 4.86 | |
| SQSD(Aegis)Backbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -4.76 | 0.23 | -1.91 | 0.27 | 3.35 | 8.11 | |
| SQSD(Aegis)Backbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | -1.25 | 2 | 3.17 | 3.78 | 5.48 | 6.73 | |
| SQSD(Beaver)Backbone=Qwen3-8B2026.05 | — | — | — | — | — | — | — | -4 | -0.62 | 1.35 | 2.53 | 3.7 | 7.7 | |
| SQSD(Beaver)Backbone=Llama3.1-8B-Instruct2026.05 | — | — | — | — | — | — | — | -4.93 | -3.64 | -2.11 | 1.63 | 1.73 | 6.66 | |
| SQSD(Beaver)Backbone=Llama2-7b-Chat2026.05 | — | — | — | — | — | — | — | -1.15 | 1.57 | 3 | 3.84 | 6 | 7.15 |